Perceptual Adaptation to Room Acoustics and Effects on Speech Intelligibility in Hearing-Impaired Populations

Summary

Recent evidence suggests that brief listening exposure to a reverberant room environment can improve closed-set speech intelligibility in that same environment. For normal-hearing populations, this room adaptation effect can result in improvements in intelligibility of as much as 20%, but depends strongly on the reverberation time of the room, and appears to require binaural input. Because poor speech intelligibility in reverberation is a common complaint for hearing-impaired listeners, it is important to determine how room adaptation might impact speech intelligibility for hearing-impaired populations. Here, room adaptation was quantified for a sample of listeners with sensorineural hearing loss that varied in severity and configuration. Speech reception thresholds (SRTs) were measured both with and without prior listening exposure to the room environment. Headphone-based auralization techniques were used to simulate the acoustics of various listening rooms, ranging from anechoic to highly reverberant space (broadband T₆₀ = 3 s). Although SRTs both with and without prior room exposure were found to be generally elevated relative to normal-hearing listeners, the room adaptation effect, as defined by the relative decrease in SRT with room exposure, was comparable on average to that observed for normal-hearing listeners. This result is consistent with the view that room adaptation effects result from central auditory processing mechanisms.

1. Introduction

Aspects of room acoustics have long been known to cause problems for speech communication. For example, increasing amounts of room reverberation are known to significantly degrade the speech signal, and these degradations can result in speech understanding deficits [1]. Recent results suggest, however, that some of this degradation may be perceptually offset when listeners are provided with prior auditory exposure to the room. Such room exposure has been shown to objectively improve speech intelligibility [2] and to modify speech perception [3, 4]. This phenomenon appears to be similar to the adaptive buildup of echo suppression observed in situations when only a single echo is present (see [5] for review). Because speech perception in reverberant rooms is known to be particularly problematic for listeners with hearing-impairment [6], it is important to determine whether hearing-impaired listeners might obtain similar benefits from prior room listening exposure as do normally-hearing listeners. This is the goal of the current study, which builds on previous work with normally-hearing listeners [2] using the same testing paradigm.

2. Methods

2.1. Listeners

A total of 26 listeners participated in this study. The hearing-impaired group was composed of 12 listeners (1 male, 11 female) ranging in age from 23 – 82 years. All had adult-onset bilaterally symmetrical (the median interaural difference in thresholds across all frequencies was 5 dB) sloping high-frequency sensorineural hearing loss. Table I displays age and pure-tone averages of airconductions thresholds at .5, 1, and 2 kHz for the left and right ear of each listener in this group. Five listeners in this group routinely used hearing aids, although all testing in this study was conducted un-aided. The normally-hearing group consisted of 14 young adults (6 male, 8 female) ranging in age from 17 – 24 years. All had pure-tone air-conduction thresholds of < 25 dB HL [7] from .25 – 8 kHz. Mean (±1 standard deviation) pure-tone air-conduction thresholds are shown in Figure 1 for both groups of listeners.

Table I.

Hearing-impaired listener information: Identification code, age (in years), pure-tone average (PTA) of air-conduction thresholds for left and right ears (see text for details), signal-to-noise ratio (SNR) testing range, supplemental signal+noise gain (see text for details), and matrix of completed (shaded) “No Carrier” (NC) and “Sentence Carrier” (SC) conditions in each test room.

Figure 1.

Mean pure-tone air-conduction audiograms for the left and right ears (filled and open symbols, respectively) for normally-hearing and hearing-impaired listeners. Shaded regions indicate 1 standard deviation about the mean.

2.2. Speech Corpus

Speech materials were taken from the Coordinate Response Measure (CRM) corpus [8].

2.3. Room Simulation

Virtual auditory space techniques were used to simulate listening conditions in five rooms, ranging from anechoic to highly-reverberant. The rooms had identical dimensions of 5.7 × 4.3 × 2.6 m, but differed in the absorptive properties of the simulated boundary surfaces. Table II displays broadband reverberation times, T₆₀, and clarity indices, C₅₀ [9], for each room. Each room was simulated using a simple model of a binaural room impulse response (BRIR), which was constructed using an image-model [10] to simulate early reflections and a statistical model of the late reverberant energy. The simulation techniques were identical to those used in a previous related study [2], and have been shown to produce results that are perceptually similar to those derived from measurements in a real room [11].

Table II.

Broadband (125 – 4000 Hz) reverberation time, T₆₀, and Clarity Index, C₅₀, measures for each test room.

Room	T60 (s)	C50 (dB)
0	<0.01	>60
1	0.31	25.8
2	0.42	13.4
3	1.2	3.5
4	3	-6.6

In all rooms, the target speech was simulated at a spatial location directly in front of the listener at a distance of 1.4 m, and a broadband noise masker was presented at a simulated location opposite the listener's right ear, also at 1.4 m. Signal-to-noise ration (SNR) varied over a range of 32 dB in 4 dB steps. For normally-hearing listeners, SNRs ranged from -28 to +4 dB. Pilot testing was used to determine an appropriate range of SNRs for each hearing-impaired listener. These ranges are displayed in Table I.

2.4. Design and Procedure

The experimental design and procedures were fundamentally identical to those used in previous work [2]. Speech reception thresholds (SRTs) were measured under two different listening exposure conditions. The “No Carrier” condition (NC) limited prior room listening exposure by presenting listeners with only the color/number targets from the CRM speech corpus and selecting the simulated room at random from trial-to-trial within a block of trials. The “Sentence Carrier” condition (SC) enhanced room exposure by presenting listeners with a two-sentence carrier phrase (~10 s duration) preceding the color/number targets and by holding the simulated room constant within a block of trials. Each block contained 54 trials (6 repetitions at each on 9 SNRs). Five trial blocks were completed for a given room/condition combination. It should be noted that not all listeners completed all room/condition combinations. Completed combinations for the hearing-impaired group are shown in Table I. The dataset is much more complete for the normally-hearing group, with the notable exception that only two listeners have completed conditions involving Room 3. Portions of the normally-hearing dataset (Room 0 and Room 2) appear in a previous publication [2].

All sounds were presented over equalized Beyerdynamic DT-990-Pro headphones using a Digital Audio Labs CardDeluxe for D/A conversion (24-bit, 44.1 kHz) within a double-walled sound isolation chamber (Acoustic Systems). For the normal hearing group, all sounds were presented at moderate levels (approximately 65 dB SPL). During pilot testing, hearing-impaired listeners were given the option to adjust the sound levels (in 5 dB steps) such that the speech signals were comfortably audible. Table I displays any supplemental gain applied to both signal and noise for individual listeners in this group relative to the nominal levels used for normally-hearing listeners. Listeners entered their responses from the CRM task on a graphical user interface. All signal processing and data collection applications were implemented using Matlab software (Mathworks, Inc.).

2.5. Data Analysis

The proportion of correct responses (PC) was computed separately for each listener in all combinations of exposure condition, room, and SNR. Logistic functions of the following form were fit to the data using a maximum likelihood procedure [12],

$$\text{Est}.PC\left(x\right)=(1-\delta )\times \frac{1}{1+\text{exp}(-(\alpha -x)∕\beta )}+\delta$$

where α is the estimated threshold parameter, β is the estimated slope parameter, and δ is chance-performance level (1/32 in the CRM task). 95% confidence intervals were obtained for each fitted function's threshold value, PC = .516, using a bootstrapping procedure [13].

3. Results

Overall, the data were well-approximated by the logistic function fits (R² > .44 in all cases, with a median R² of .97). Figure 2 displays representative function fits for one listener (LMN) for each of the experimental conditions in Room 1. The room adaptation effect may be observed in Figure 2 as a decrease in threshold between the NC and SC conditions. This is presumably due to a buildup of echo suppression. Slopes of the fitted functions in Figure 2 are relatively homogeneous, and this is representative of the fits in general across listeners, conditions, and rooms. Estimated psychometric functions are therefore well-described by their threshold parameters alone.

Figure 2.

Proportion of correct speech target identifications as a function of signal-to-noise ratio and estimated psychometric functions for a single listener (LMN) in Room 1 for both the No Carrier (NC) and Sentence Carrier (SC) experimental conditions. Speech reception threshold (SRT) estimates (with 95% confidence intervals) are displayed for each fitted function. Speech intelligibility enhancement with prior listening exposure (“buildup”) is indicated by decreased SC SRT relative to the NC SRT

Figure 3 displays speech reception thresholds in all rooms for the NC condition. Results for both normally-hearing and hearing-impaired listeners are shown. As expected, generally elevated SRTs are observed for the hearing-impaired listeners relative to normally hearing listeners. In addition, SRTs increase with increasing reverberation for both groups of listeners, consistent with previous work related to the effects of room reverberation on speech intelligibility [14].

Figure 3.

Summary of SRTs in the NC condition in all rooms for all listeners. SRTs for the normally-hearing group are indicated by black symbols. SRTs for the hearing-impaired group are indicated by red symbols. Subject identification codes are also indicated for all hearing-impaired listeners. Bars show 95% confidence intervals about each SRT

Figure 4 plots the difference in SRT between the NC and SC conditions. Positive values indicate a reduction in SRT (improved performance) in the SC condition relative to the NC condition. Bars indicate 95% confidence limits for each difference, estimated using information from the bootstrapped confidence limits about individual-listener SRTs in each condition (see Figure 2 for examples). Determining these confidence limits is complicated by the fact that estimates of a given listener's thresholds are likely not independent across different experimental conditions. We therefore computed confidence limits using the following relationship for the variance of the difference between two random variables, a and b:

$$\text{var}\left(\mathrm{a}\text{-}\mathrm{b}\right)=\text{var}\left(\mathrm{a}\right)+\text{var}\left(\mathrm{b}\right)\text{-}2r\sqrt{\text{var}\left(\mathrm{a}\right)}\sqrt{\text{var}\left(\mathrm{b}\right)}$$

where r is the Pearson correlation between a and b. Confidence limits in Figure 4 were computed based on a value of r = 0.7.

Figure 4.

Difference in SRT values between the NC and SC conditions in all rooms. Data for both normally-hearing (black symbols) and hearing-impaired (red symbols) are shown. Positive values indicate lower SRT (better performance) in the SC condition relative to the NC condition. Bars show 95% confidence intervals about each SRT difference (see text for details). Subject identification codes are indicated for all hearing-impaired listeners.

At least two noteworthy patterns are observable in the data displayed in Figure 4. First, the size of the adaptation effect appears to depend on the simulated room. In general, consistent adaptation effects are not observed in anechoic space (Room 0) or highly reverberant space (Room 4), but are increasingly observed as room reverberation is increased over an intermediate range (Rooms 1 – 3). These are important results, because they suggest that the effect is linked specifically to the acoustical properties of the room (since the effects are not observed without a room, i.e. in anechoic space) and that the effect is strongest in moderately reverberant rooms. Similar results have been reported in related studies from our laboratory [15]. For reference, the average SRT improvement in Room 2 was 2.7 dB for the normally-hearing listeners. This corresponds to a greater than 18% improvement in speech understanding in the SC condition relative to threshold SNR for the NC condition [2].

A second, and perhaps more practically important pattern observable in Figure 4 is that the effect sizes do not appear to be markedly different between the normally-hearing and hearing-impaired groups of listeners. This pattern was confirmed by pooling the data across rooms, and comparing mean effect sizes from each group of listeners. No statistically significant difference in effect sizes was found, t(95) = -.389, p = .698. Although increases in the effect size variability can be observed in certain rooms (e.g. Room 2) for the hearing-impaired group, overall the similarity in effect sizes is surprising, given the large differences in NC SRTs between the two groups (Figure 3).

This similarity in adaptation effects between normally-hearing and hearing-imparied groups is important, because it appears to minimize the role of the auditory periphery such effects. It is also consistent with other studies related aspects of reflective sound processing that are thought to be mediated by centralized brain mechanisms [16]. Although the specifics of these mechanisms have yet to be identified, it has been suggested that the mechanisms involve high-level perceptual calibration to particular aspects of the listening environment's acoustics [17, 18], perhaps related to its modulation transfer function [19]. Regardless of cause, the fact that hearing impaired listeners show similar improvements in speech understanding in reverberant rooms when allowed brief periods of prior listening exposure may have important practical implications for aural rehabilitation.

4. Conclusions

Hearing-impaired listeners appear to derive roughly similar benefits from prior listening exposure to a room as do normally-hearing listeners. Brief periods (on the order of seconds) of exposure result in improvements in speech reception threshold of 2 to 3 dB, which correspond to improvements in speech intelligibility of nearly 20%. These improvements appear to be specific to rooms (they are not consistently observed in anechoic space) and are strongest for rooms with moderate amounts of reverberation (0.3 < T₆₀ < 1.2 s).